SSD Caching: Device-Mapper- and Hardware-based solutions compared Werner Fischer & Georg Schönberger Technology Specialists Thomas-Krenn.AG 18. LinuxTag / Berlin / Germany 24th May 2011 Agenda • Introduction • Technologies compared • A few details • Performance results • Test strategy • Lessons learned slide 2/45 Introduction Source: xkcd.com Do things the right way! slide 3/45 Introduction who is Server & accessories based in Freyung, serving all over Europe "Made in Germany" Bavaria slide 4/45 Should I use SSD caching? slide 5/45 Well, Yes and No... if(dataSet == known && ramAvailable != enough && appAccess == analyzed && perfTests == available) CheckForCacheTechnologies(); else NeedMoreInfo(); slide 6/45 Establish some wording... slide 7/45 Caching Basics slide 8/45 Some main features compared... slide 9/45 SSD Caching – Compared 1 of 6 WB WT WA/read-only FlashCache CacheCade 1 MaxCache ! 2 1Including ForcedWB ! Attention: No redundant cache with multiple SSDs possible 2Including InstantWB slide 10/45 SSD Caching – Comparison 2 of 6 Skip sequential I/O FlashCache 1 CacheCade 2 MaxCache 3 • Maybe your disk RAID is faster for sequential I/O 1Configurable via sysctl (threshold) 2No further details known 3Not configurable, always skipped slide 11/45 SSD Caching – Comparison 3 of 6 Lazy Cleaning FlashCache 1 CacheCade 2 MaxCache 3 • WB: clean dirty pages from SSD to HDD 1Based on threshold and idleness – Dynamic via sysctl 2No further details known 3Background flush with dynamic rate – FlushAndFetchRate – dirtyPageThreshold slide 12/45 SSD Caching – Comparison 4 of 6 Cache persistence FlashCache ! 1 CacheCade 2 MaxCache 3 1Only WB Cache is persistent – https://groups.google.com/forum/#!topic/flashcache-dev/xRW64FOxgWw – People are discussing to add persistence for WT 1On node crash, only Dirty Blocks are in cache – https://groups.google.com/forum/#!topic/flashcache-dev/LqjKZcGL1eU 2Persistent for all cache modes 3Dirty shutdown: WT – discarded, WB – LV failed slide 13/45 SSD Caching – Comparison 5 of 6 Error WB WT strategy FlashCache 1 Currently - Error2 CacheCade 3 MaxCache 4 5 1No switch from WB to WT if SSD redundant 2Patch to make WB/WT errors transparent – https://groups.google.com/forum/#!topic/flashcache-dev/6XEhY87uv1g 3Details on slide 26 4With write cache LV fails 5No further information available slide 14/45 SSD Caching – Comparison 6 of 6 Hot spot detection FlashCache CacheCade MaxCache slide 15/45 FlashCache CacheCade MaxCache Take a closer look... slide 17/45 Not an official logo. Source: deviantart.com FlashCache Not all about Facebook is bad... Visit https://github.com/facebook/flashcache/ slide 18/45 FlashCache – Introduction • Open Source :D – Get code from Github • Using Linux Device Mapper (cf. LVM) – Visit our I/O stack diagram • http://www.thomas-krenn.com/en/oss/linux-io-stack-diagram.html – Based on dm-cache • Easy to deploy and use – Compile – Load one kernel module slide 19/45 FlashCache – Highlights • FIFO/LRU replacement policy – online changeable • Define process blacklist/whitelist • Cleaning of dirty blocks configurable – Threshold – Idleness • Error injection flags via sysctl – E.g. READCACHE_ERROR • Detailed Cache statistics – dmsetup table slide 21/45 FlashCache – Statistics fc-root: 0 3907029168 flashcache conf: ssd dev (/dev/sdd), disk dev (/dev/sdc) cache mode(WRITE_BACK) capacity(32638M), associativity(512), data block size(4K) metadata block size(4096b) skip sequential thresh(0K) total blocks(8355328), cached blocks(1048668), cache percent(12) dirty blocks(1048662), dirty percent(12) nr_queued(0) Size Hist: 1024:1 4096:9765813 512:0 1024:1 1536:0 2048:0 2560:0 3072:0 3584:0 4096:9765813 4608:0 5120:0 5632:0 6144:0 6656:0 7168:0 7680:0 8192:0 8704:0 9216:0 9728:0 10240:0 10752:0 11264:0 11776:0 12288:0 12800:0 13312:0 13824:0 14336:0 14848:0 15360:0 15872:0 16384:0 reads=1048896 writes=8716916 read_hits=1048580 read_hit_percent=99 write_hits=9 write_hit_percent=0 dirty_write_hits=0 dirty_write_hit_percent=0 replacement=0 write_replacement=0 write_invalidates=80 read_invalidates=0 pending_enqueues=0 pending_inval=0 metadata_dirties=1048662 metadata_cleans=0 metadata_batch=1042405 metadata_ssd_writes=6257 cleanings=0 fallow_cleanings=0 no_room=0 front_merge=0 back_merge=0 disk_reads=316 disk_writes=7668261 ssd_reads=1048580 ssd_writes=1055014 uncached_reads=221 uncached_writes=7668261 uncached_IO_requeue=0 uncached_sequential_reads=0 uncached_sequential_writes=0 pid_adds=0 pid_dels=0 pid_drops=0 pid_expiry=0 slide 22/45 FlashCache – Considerations • Make your cache redundant (WB) – What if SSD fails? – Cf. https://groups.google.com/forum/#!topic/flashcache-dev/6XEhY87uv1g • With WB partial writes are possible – There are ideas to fix this (cf. flashcache-doc.txt) • Only WB cache is persistent – Across reboots and cache removals – No metadata on SSD for WT/WA • Read https://groups.google.com/forum/#!topic/flashcache-dev/xRW64FOxgWw slide 23/45 LSI MegaRAID CacheCade slide 24/45 CacheCade – Introduction • Hardware-based solution • Caching algorithm in controller • Redundant SSD caches possible – Avoid data loss with WB slide 25/45 CacheCade – Error handling WB WT FWB Red. Change to Read SSD WB WT Non. No VD Read HDD No VD access Red. access • Except WB and non redundant SSD Cache – Application doesn't notice SSD failure • Controller reports over Storage Manager slide 26/45 CacheCade – Considerations • Our tests showed that... – Assign → initialization – Fio laying out files • Not only on SSD (cf. FlashCache) • Background syncs – Test Read values not correct – Fio with 'create_only' • 100 MB/s read • Data not completely in cache • → read a 2nd time → 215MB/s slide 27/45 Adaptec MaxCache slide 28/45 MaxCache – Features • Manual – http://download.adaptec.com/pdfs/user_guides/cli_v7_30_18837_users_guide.pdf • WriteCache – Should not be used until now – Redundant cache cannot be created • Sequential I/O is not cached – SAS drives may be faster – Performance stays at 110 MB/s with streaming I/O • Hot spot detection logic slide 29/45 At the end its all down to performance... Prepare to get this... People might promise you... Test your own specific application. If you want with a little help from us :) slide 33/45 Comparison – Bandwidth read slide 34/45 Comparison – Bandwidth write slide 35/45 Comparison – Bandwidth randread slide 36/45 Comparison – Bandwidth randwrite slide 37/45 Still to analyze... slide 38/45 FlashCache ? slide 39/45 CacheCade ? slide 40/45 Summary – avoiding errors while testing... slide 41/45 SSDs – A well known error slide 42/45 Testing caveats slide 43/45 Lessons Learned Establish a Baseline • Test your own application Know application – Get a comfortable testing tool • I prefer Fio (git://git.kernel.dk/fio.git) Yes RAM – Know how and what to test • Check your test results twice No HDD is – Visualize your results bottleneck – Test numbers against the real world No SSD alone • For a first attempt to small – Test with FlashCache Test caching WB WT slide 44/45 Questions? Do not hesitate to ask us (you can do it per mail also) You want some test protocols? You need a special chart? slide 45/45 Thanks for your time Visit our booth - Open Source Sponsorship and Low Energy Server References • http://greenarrow.deviantart.com/art/The-Flash-Wally-West-35885287 • https://upload.wikimedia.org/wikipedia/commons/6/62/2006-07-26AbendhimmelRemstal19.jpg • http://openclipart.org/detail/15238/tortue1-by-roym • http://openclipart.org/detail/34657/architetto---struzzo-by-anonymous • http://openclipart.org/detail/11118/check-sign-and-cross-sign-by-jetxee • http://openclipart.org/detail/124837/led-lamp-by-eggib • http://b--art.deviantart.com/art/Riddler-62371451 slide 47/45 Backup Slides slide 48/45 Test system • SSDs – Intel Series 320 160GB – Via HPA reduced to 32GB • RAID Controller – LSI MegaRAID SAS 9260-4i – Adaptec 6805Q • Software – Fio 2.0.7 – Ubuntu 12.04 • Updates from Release Day slide 49/45 Flashcache – Test Script • flashcache_create – Create a WB caching device • cache_all=0 – Don't cache ext4 initialization • mkfs.ext4 -q -E lazy_itable_init=0, lazy_journal_init=0 /dev/mapper/fc-root • mount /dev/mapper/fc-root • cache_all=1 • Call fio • umount /dev/mapper/fc-root • dmsetup remove • flashcache_destroy /dev/sdd slide 50/45 Flashcache – Workflow slide 51/45 Flashcache – Commands • Just a few examples – flashcache_create -v -p back fc-root /dev/sdd /dev/sdc – sysctl -w dev.flashcache.sdd+sdc.cache_all=0 – dmsetup table – flashcache_destroy /dev/sdd slide 52/45.